Stacked Deconvolutional Network for Semantic Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Image Analysis

Background:

• Semantic segmentation research has focused on enhancing spatial resolution within Fully Convolutional Networks (FCNs).
• Existing methods face challenges in effectively integrating contextual information and recovering precise localization details.

Purpose of the Study:

• To introduce a novel Stacked Deconvolutional Network (SDN) architecture for advancing semantic segmentation.
• To improve the integration of contextual information and the recovery of fine-grained localization details in segmented images.

Main Methods:

• Proposing a Stacked Deconvolutional Network (SDN) composed of multiple stacked shallow deconvolutional networks (SDN units).
• Incorporating inter-unit and intra-unit connections to facilitate information and gradient flow, aiding network training and feature fusion.
• Implementing hierarchical supervision during the upsampling process within each SDN unit to enhance feature discrimination and network optimization.

Main Results:

• Achieving new state-of-the-art performance on four benchmark datasets: PASCAL VOC 2012, CamVid, GATECH, and COCO Stuff.
• Demonstrating a top performance with an 86.6% intersection-over-union (IoU) score on the test set, even without Conditional Random Field (CRF) post-processing.

Conclusions:

• The proposed Stacked Deconvolutional Network (SDN) effectively addresses limitations in spatial resolution and localization for semantic segmentation.
• The architectural innovations, including stacked units, enhanced connections, and hierarchical supervision, contribute to superior performance and network optimization.